Short-term load forecasting (STLF) is a conventional process at power companies to serve for better decision-making in their daily operations. Weather factors play a key role in STLF. In practice, an online STLF system typically requires the use of weather forecasts as input when projecting the future load, with associated weather forecast errors. This type of forecasting is known as ex-ante forecasting. Nevertheless, most existing academic literature developed load forecasting techniques under the ex-post forecasting settings, where the actual weather information is used in the forecast period. Meanwhile, the robustness of STLF models to the real weather forecast errors has rarely been studied in the literature. The gap between the practice and the research study is often due to the shortage of historical weather forecasts. In this research, we aim to close this gap by proposing two new frameworks to select better models in short-term ex-ante load forecasting. Compared to the conventional research which focuses on ex-post load forecast accuracy in the model development, both frameworks consider the impact of real weather forecast errors and are better fitted to field practices.
The effectiveness of the proposed frameworks is confirmed using an empirical case study at a medium-sized US utility with load data from multiple supply areas and real temperature forecasts. Compared to a state-of-the-art benchmark that uses the historical ex-post load forecast accuracy for model selection, the first framework leads to 2.4% improved accuracy on average. A further study among the weather sensitive hours (i.e., the hours when a smaller error in the temperature forecast may lead to a greater inaccuracy in the load forecast) suggests that the first framework outperforms the benchmark by 3.1% on average, although its performance is subpar when the predicted temperature forecast accuracy gets worse. The second framework addresses this issue effectively and improves the accuracy of the first framework by 7.4% for the hours with worse predicted temperature forecast accuracy. Overall, the second framework leads to an average of 0.8% improvement over the first framework and 3.9% improvement over the benchmark among the weather sensitive hours.